10
AbstractIn signal processing, important information is mainly required and processed. Discrete Cosine Transform (DCT) provides a great compaction capabilities. One of the challenges of face recognition using DCT and any other algorithm is poor illumination of the acquired images. In this paper, anisotropic diffusion illumination normalization technique (AS) and DCT were used for recognition. The AS was employed as a preprocessor before applying the DCT as a feature extractor. The new face recognition technique named ‘ASDCT’ was assessed on ORL, Yale, and extended Yale-B databases. Performance metrics were generated and evaluated by verification and identification rates using nearest neighbor classifier (NNC). Appearance based techniques were also exploited for comparison with the new method, and results show that ‘ASDCT’ outperformed many renowned algorithms in the literature. It has produced up to 93.4% verification rate at False Acceptance Rate (FAR) =0.1% on ORL database. Therefore, its performance using both controlled and uncontrolled databases is considerably good. It is believed that, the new framework enhances the DCT feature extraction for more efficient face recognition. Index TermsDecorrelation, discrete cosine transform, face recognition, illumination, nearest neighbor classifier I. INTRODUCTION IOMETRICS is essentially a pattern recognition system [1], which uses physiological and behavioral characteristics of a person. Physiological characteristics may be fingerprint, eye (iris-pattern, retina-scan), face, hand geometry or palm-print [2]. Whereas behavioral characteristics may be voice, signature, or other keystrokes dynamics. Such characteristics are called ‘traits’, they are unique and varied across persons. They are found to be general and natural, user-friendly and nonintrusive. Manuscript received February 28, 2015; revised October 28, 2015. This work was financially supported by Kano State Scholarship Board, batch 502, under the Kano State Government Nigeria, and academically supported by Universiti Sultan Zainal Abidin (UniSZA), Kuala Terengganu, Terengganu, Malaysia. Z. Sufyanu was an Electrical Engineer in Dangote Cement Plant Obajana, Lokoja, Kogi State, Nigeria. He is at present with Faculty of Informatics and Computing, Universiti Sultan Zainal Abidin Malaysia (email: [email protected]). F. S. Mohamad is a Deputy Dean of Graduate School, Universiti Sultan Zainal Abidin Malaysia (phone: +601-9906-6074; e-mail: [email protected]). A. A. Yusuf is with the Faculty of Informatics and Computing, Universiti Sultan Zainal Abidin, Malaysia, on leave from National Biotechnology Development Agency (NABDA), Abuja, Nigeria (e-mail: [email protected]). M. B. Mamat is currently a Professor at Universiti Sultan Zainal Abidin Malaysia (e-mail: [email protected]). However, the extensive variation of those features over one another makes them attractive to the researchers. Besides, they are personal; one cannot forget or lose them except in an accident. The history of face recognition technology started since 1960s with the development of a semi- automated system [9]. Thereafter, massive advancement has been made in this field. Face recognition system was not the first biometric technology to be introduced, but has become one of the most attending biometric attributes. It is also the most important biometric authentication technology. The advent of face recognition solves more advanced problems of authentication in highly secured systems. Thus, automatic face recognition has received much attention within the computer vision community over the past three decades [3]. And it spans numerous fields and disciplines [4]. Although there are other reliable methods of biometric identification such as fingerprint and iris scans, but face recognition is proven more effective. The advantages of facial identification lie on cost and user’s convenience. Six biometric attributes were considered by Hietmeyer in [5], where facial features scored highest in terms of compatibility in a Machine Readable Travel Documents (MRTD) system. Therefore, commercial application of face technology has received large interest, especially in credit card companies to reduce fraudulent usage of the credit cards. Hence, encoding each card with facial features for identification would highly minimize bank card frauds. In addition, human’s face recognition is a matured field which resulted to a higher demand for accurate and fast user authentication. Although this field is considered matured for automatic solutions, there are always problems right from image acquisition to classification. And researchers considered these problems very challenging to solve. Meanwhile, to attain a successful recognition performance, most current recognition approaches require some control over the imaging conditions, because many real-world applications require operational flexibility [6]. The proper recognition under pose variations, illuminations, occlusions and expressions is the optimal target of any algorithm and solution against any unauthorised action. Many research focus on establishing secured identification systems. A face system is separated into three main stages namely; preprocessing, feature extraction, classification and recognition [7]. Furthermore, several improvements on recognition systems were made, yet there was no benchmark strategy for recognition systems [8]. Researches on face recognition mostly fall into two main categories, namely: holistic and feature-based [9]. Holistic based is the global approach to face recognition which relies on encoding the entire facial Enhanced Face Recognition Using Discrete Cosine Transform Zahraddeen Sufyanu, Member, IAENG, Fatma S. Mohamad, Abdulganiyu A. Yusuf, and Mustafa B. Mamat B Engineering Letters, 24:1, EL_24_1_08 (Advance online publication: 29 February 2016) ______________________________________________________________________________________

Enhanced Face Recognition Using Discrete Cosine Transform · Discrete Cosine Transform (DCT) provides a great compaction capabilities. One of ... be fingerprint, eye (iris-pattern,

  • Upload
    hathuan

  • View
    227

  • Download
    0

Embed Size (px)

Citation preview

Abstract— In signal processing, important information is

mainly required and processed. Discrete Cosine Transform

(DCT) provides a great compaction capabilities. One of the

challenges of face recognition using DCT and any other

algorithm is poor illumination of the acquired images. In this

paper, anisotropic diffusion illumination normalization

technique (AS) and DCT were used for recognition. The AS

was employed as a preprocessor before applying the DCT as a

feature extractor. The new face recognition technique named

‘ASDCT’ was assessed on ORL, Yale, and extended Yale-B

databases. Performance metrics were generated and evaluated

by verification and identification rates using nearest neighbor

classifier (NNC). Appearance based techniques were also

exploited for comparison with the new method, and results

show that ‘ASDCT’ outperformed many renowned algorithms

in the literature. It has produced up to 93.4% verification rate

at False Acceptance Rate (FAR) =0.1% on ORL database.

Therefore, its performance using both controlled and

uncontrolled databases is considerably good. It is believed that,

the new framework enhances the DCT feature extraction for

more efficient face recognition. Index Terms— Decorrelation, discrete cosine transform, face

recognition, illumination, nearest neighbor classifier

I. INTRODUCTION

IOMETRICS is essentially a pattern recognition system

[1], which uses physiological and behavioral

characteristics of a person. Physiological characteristics may

be fingerprint, eye (iris-pattern, retina-scan), face, hand

geometry or palm-print [2]. Whereas behavioral

characteristics may be voice, signature, or other keystrokes

dynamics. Such characteristics are called ‘traits’, they are

unique and varied across persons. They are found to be

general and natural, user-friendly and nonintrusive.

Manuscript received February 28, 2015; revised October 28, 2015. This

work was financially supported by Kano State Scholarship Board, batch

502, under the Kano State Government Nigeria, and academically

supported by Universiti Sultan Zainal Abidin (UniSZA), Kuala

Terengganu, Terengganu, Malaysia.

Z. Sufyanu was an Electrical Engineer in Dangote Cement Plant

Obajana, Lokoja, Kogi State, Nigeria. He is at present with Faculty of

Informatics and Computing, Universiti Sultan Zainal Abidin Malaysia

(email: [email protected]).

F. S. Mohamad is a Deputy Dean of Graduate School, Universiti Sultan

Zainal Abidin Malaysia (phone: +601-9906-6074; e-mail:

[email protected]).

A. A. Yusuf is with the Faculty of Informatics and Computing,

Universiti Sultan Zainal Abidin, Malaysia, on leave from National

Biotechnology Development Agency (NABDA), Abuja, Nigeria (e-mail:

[email protected]).

M. B. Mamat is currently a Professor at Universiti Sultan Zainal Abidin

Malaysia (e-mail: [email protected]).

However, the extensive variation of those features over one

another makes them attractive to the researchers. Besides,

they are personal; one cannot forget or lose them except in

an accident. The history of face recognition technology

started since 1960s with the development of a semi-

automated system [9]. Thereafter, massive advancement has

been made in this field. Face recognition system was not the

first biometric technology to be introduced, but has become

one of the most attending biometric attributes. It is also the

most important biometric authentication technology. The

advent of face recognition solves more advanced problems

of authentication in highly secured systems. Thus, automatic

face recognition has received much attention within the

computer vision community over the past three decades [3].

And it spans numerous fields and disciplines [4].

Although there are other reliable methods of biometric

identification such as fingerprint and iris scans, but face

recognition is proven more effective. The advantages of

facial identification lie on cost and user’s convenience. Six

biometric attributes were considered by Hietmeyer in [5],

where facial features scored highest in terms of compatibility

in a Machine Readable Travel Documents (MRTD) system.

Therefore, commercial application of face technology has

received large interest, especially in credit card companies to

reduce fraudulent usage of the credit cards. Hence, encoding

each card with facial features for identification would highly

minimize bank card frauds.

In addition, human’s face recognition is a matured field

which resulted to a higher demand for accurate and fast user

authentication. Although this field is considered matured for

automatic solutions, there are always problems right from

image acquisition to classification. And researchers

considered these problems very challenging to solve.

Meanwhile, to attain a successful recognition performance,

most current recognition approaches require some control

over the imaging conditions, because many real-world

applications require operational flexibility [6]. The proper

recognition under pose variations, illuminations, occlusions

and expressions is the optimal target of any algorithm and

solution against any unauthorised action. Many research

focus on establishing secured identification systems. A face

system is separated into three main stages namely;

preprocessing, feature extraction, classification and

recognition [7].

Furthermore, several improvements on recognition

systems were made, yet there was no benchmark strategy for

recognition systems [8]. Researches on face recognition

mostly fall into two main categories, namely: holistic and

feature-based [9]. Holistic based is the global approach to

face recognition which relies on encoding the entire facial

Enhanced Face Recognition Using Discrete

Cosine Transform

Zahraddeen Sufyanu, Member, IAENG, Fatma S. Mohamad, Abdulganiyu A. Yusuf, and Mustafa B.

Mamat

B

Engineering Letters, 24:1, EL_24_1_08

(Advance online publication: 29 February 2016)

______________________________________________________________________________________

image, whereas feature-based approach to face recognition

relies on detection and characterization of individual facial

features [10]. Ahmed and Rao in [11] initially introduced a

Discrete Cosine Transform (DCT) in the early seventies.

Since then, the DCT has become very popular and several

versions of it have been proposed [12]. Few numbers of

DCT coefficients were used to reduce redundancy and

recover the original image from the selected coefficients.

The coefficients with largest magnitude (i.e. low frequency)

are mainly located in the upper left corner of the DCT

matrix. This section is related to illumination variation and

smooth regions such as forehead and cheeks of the face

image. But the coefficients with lowest magnitude (i.e. high

frequency) are situated in the bottom right corner of the

DCT matrix. This section is related to noise and detailed

information about edges in the image. And the coefficients

with mid magnitude (i.e. medium frequency) are found in the

middle region. This represents the general structure of the

face in the image.

The three sections appear in DCT matrix as represented in

Fig. 1. Therefore, amplitude distribution of the DCT

coefficients for 8 x 8 blocks measured for a test image is

depicted in Fig. 2. This shows the variations of DCT

coefficients of an image.

Fig. 1. Three regions of DCT coefficient matrix

Fig. 2. Histogram of DCT coefficients of a ‘bridge’ image

Performance of DCT is largely affected by altering the

coefficients magnitude at the top left corner of the matrix.

Thus, illumination normalization is applied prior to feature

extraction to compensate these coefficient variances and set

the blocks to more equalized intensity.

The DCT is an efficient technique for image coding, and

has been successfully used for face recognition in [17], [18],

[19], and [20]. Since the most high frequency coefficients of

DCT are almost zero, such coefficients can be ignored

without degrading the image quality. In addition, different

coefficients can be quantized based on visual sensitivity with

different accuracy.

Many researchers have tried to truncate the number of

DCT coefficients. The research conducted by Bilal in [13]

suggested saving the first five values from every DCT block

to restore the image without significant errors. In an attempt

to increase the performance of a face system with variation

in facial geometry and illumination [14], 2D-DCT method

was proposed for feature extraction. The DCT coefficients

of face images were truncated in an exponential way, and

then a Principal Component Analysis (PCA) was applied for

dimensionality reduction. Only a few non-zero eigenvalues

related eigenvectors were considered and verified using K-

nearest distance measurement. Sanderson and Paliwal in

[15] suggested a feature extraction technique called DCT-

mod2. The DCT was first applied to sub regions of facial

images to extract a number of DCT coefficients. Then any

potential illumination change was compensated through

replacing the coefficients most affected by illumination with

their corresponding horizontal and vertical delta coefficients.

This derived a face representation insensitive to external

lighting changes. And various databases with illumination-

induced variability were used to assess this method. The

DCT-mod2 technique resulted in promising results on all

databases considered. Recently, an advanced DCT through

histograms was developed in [16], which improves the

compression ability of the DCT.

The feature extraction under uncontrolled illumination

conditions is a major concern and a great challenge for

recognition in real world applications [8]. Illumination

normalization (IN) is a preprocessing technique that

compensates for different lighting conditions. However,

regardless of a number of IN techniques in the literature, the

choice of a technique that improves a particular algorithm no

matter how little is very essential. Researchers were

attempted to improve face recognition under uncontrolled

lighting conditions, through the choice of various techniques

for illumination compensation and preprocessing

enhancement.

Before carrying out the feature extraction using variety of

methods, different illumination techniques were used to

normalize the illuminations under such constraints. This is to

restore a face image back to its normal lighting condition

[21], [22], [23], [24], [25], [26]. To the best of the authors’

knowledge, none has previously focussed on face

recognition using DCT by finding the most suitable

preprocessor. Our main motivation behind emphasizing on

the DCT is because; most practical transform coding systems

are based on DCT, since it provides a good compromise

between information packing ability and computational

complexity [27].

The present study focuses on preprocessing and feature

extraction to tackle illumination challenges and improve

decorrelation power of DCT for more efficient face

recognition. The research also considers recognition under

various expressions.

Low

Medium

High

Engineering Letters, 24:1, EL_24_1_08

(Advance online publication: 29 February 2016)

______________________________________________________________________________________

Section II of the paper describes evaluation process of the

preprocessing techniques, and section III analyzes the

experimental setups. In section IV, results of the proposed

methodology and other techniques are assessed. Section V

exposes the effects of truncation in DCT domain, whereas in

section VI concluding remarks and future study are outlined.

II. PREPROCESSING

A. Introduction

Normalizing an illumination from images before extracting

their features using DCT, compensates the coefficient

variances and sets the blocks to more equalized intensity. It

is very important to categorize preprocessors according to

their suitability on a particular algorithm. For this reason, an

empirical study was conducted to choose the best IN

technique specifically for DCT performance improvement.

The established 22 IN techniques were tested using extended

Yale-B database. The IN technique that will be subjected to

compression equivalent to that of DCT and be reconstructed

with least error is the goal of this evaluation.

B. Illumination Normalization (IN) Techniques

The well-known 22 IN techniques proposed in the

literature are:

(1) single scale retinex (SSR),

(2) multi scale retinex (MSR),

(3) adaptive single scale retinex (ASR),

(4) homomorphic filtering (HOMO),

(5) single scale self quotient image (SSQ),

(6) multi scale self quotient image (MSQ),

(7) discrete cosine transform (DCT),

(8) retina modeling (RET),

(9) wavelet (WA),

(10) wavelet denoising (WD),

(11) isotropic diffusion (IS),

(12) anisotropic diffusion (AS),

(13) steering filter (SF),

(14) non-local means (NLM),

(15) adaptive non-local means (ANL),

(16) modified anisotropic diffusion (MAS),

(17) gradientfaces (GRF),

(18) single scale weberfaces (WEB),

(19) multiscale weberfaces (MSW),

(20) large and small scale features (LSSF),

(21) tan and triggs (TT), and

(22) difference of gaussian filtering (DOG).

Fig. 3 describes results of the different IN techniques

applied on few images from extended Yale-B database.

SSR MSR ASR HOMO WD

SSQ DCT MSQ SF IS

WA AS NLM MAS WEB

ANL MSW DOG GRF TT

LSSF RET

Fig. 3. Results showing effect of the 22 IN techniques applied on

extended Yale-B images

C. Evaluation of the IN Techniques

Several images from the extended Yale-B were

automatically normalized one after the other by

implementing each of the IN technique source codes. These

images were compressed using principle of JPEG

compression. The compression was carried out to test which

illumination technique would compensate lighting effect

without significantly degrading the quality of the images.

Accordingly, the following steps were computed:

1) Input image

2) Set compression quality between 1 to 100

3) Apply compression using DCT

4) Reconstruct the images

4) Plot both the original image and the DCT

compressed image

5) Save the images to file

6) Perform error calculations

For compression in this study, 60% of the coefficients

were extracted. The compressed features were obtained by

run-length encoding, and then reconstructed using inverse

DCT. The reconstructed features were retrieved to determine

the quality variation between each original image of IN

technique and its reconstructed image. After that, error

measurements were computed using these metrics: Peak

Signal to Noise Ratio (PSNR), Mean Square Error (MSE),

Root Mean Square Error (RMSE), and Mean Absolute Error

(MAE). In addition, average results of the images were

calculated.

The results reported AS technique as the highest PSNR

and the lowest MAE as indicated in Table I. This signifies

its best illumination compensation ability. For this reason,

the AS is proven as the best and most suitable technique for

recognition using the DCT. Hence, the filtered images of AS

demonstrated high similarity with their original images

despite lossy compression. The AS was introduced to the

field of face recognition by Gross and Brajovic in [28]. The

technique estimates the luminance function using anisotropic

smoothing. Therefore, the design of DCT with the AS will

compensate lighting effects, noise, and preserve object

boundary effectively.

Engineering Letters, 24:1, EL_24_1_08

(Advance online publication: 29 February 2016)

______________________________________________________________________________________

TABLE I COMPARISONs OF THE 22 ILLUMINATION NORMALIZATION TECHNIQUES, ONLY

PSNR HAS UNIT OF DB

III. EXPERIMENTAL SETUP

This section describes the experimental setup for the

study. The DCT was applied on the preprocessed images

implemented using the AS technique. The performance of

the proposed DCT and other techniques were evaluated on

Olivetti Research Laboratory (ORL), Yale and extended

Yale-B face databases. ORL and Yale are well-known and

most commonly used databases for face recognition using

the DCT in the literature. These databases provide a

successful accuracy. The ORL contains 10 different images

of 40 distinct subjects. Each subject involves images of the

following two orientations: slightly lighting variations and

facial expressions which comprise open or closed eyes,

smiling or non-smiling and with or without glasses. All the

images were taken against a dark homogeneous background

and the subjects are in up-right, frontal position with

tolerance for some side movement. The Yale face database

contains 165 face images of 15 individuals. There are 11

images per subject; each subject involves images of the

following facial expression or configuration: center-light,

glasses/no glasses, happy, left-light, normal, right-light, sad,

sleepy, surprised, and wink. The extended Yale-B was also

chosen to evaluate the recognition despite illumination

variations. This database contains gray scale frontal images

of 38 subjects under 9 poses and 64 illumination conditions.

The entire test set images used in the experiment were

manually aligned, cropped, and then resized to 168 x 192.

[29]. Samples of images in these databases are shown in

Fig. 4. For the ORL database all images were used for the

study, whereas in Yale and extended Yale-B face databases

10 images from each subject were used, leading to a total of

150 images and 280 images respectively. In addition, the

selection was done randomly for the extended Yale-B.

Fig. 4. Sample of images extracted from ORL (first row), Yale (second

row) and extended Yale-B (third row) databases

After the necessary preprocessing to make the images

suitable for efficient feature extraction, an AS was applied to

remove the illumination variations. In all the experiments,

training was performed using three images from each user in

the databases and seven others were used for subsequent

testing. Sample images of one subject used for training and

testing is shown in Fig. 5.

Fig. 5. Example of training images (first row) and testing images (second

row) from a subject in ORL database

Suppose is the number of training images, and is the

total number of images belonging to a particular subject in

each database. The training process begins by computing a

mean image from the training images. Therefore, the mean

image can be described in (1).

where represents a sample image of subject

from the number of subjects considered.

Then the DCT of the mean image is computed.

The definition of the DCT for an input image and

output image of size can be written in (2).

However, 19 coefficients were retained from each DCT

matrix and thereby used for recognition.

S/N TECHNIQUE PSNR MSE RMSE MAE

1 WEB 53.885 0.804 0.897 1.483

2 WD 53.003 0.985 0.992 1.687

3 WA 53.650 0.849 0.921 1.368

4 TT 53.912 0.799 0.894 1.498

5 SSR 52.629 1.073 1.036 1.752

6 SSQ 51.114 1.521 1.233 2.052

7 SF 55.830 0.514 0.717 1.232

8 RET 52.425 1.125 1.061 1.800

9 NLM 51.760 1.311 1.145 1.946

10 MSW 53.060 0.972 0.986 1.650

11 MSR 51.990 1.243 1.115 1.876

12 MSQ 50.171 1.890 1.375 2.297

13 MAS 54.348 0.722 0.850 1.432

14 LSSF 54.937 0.631 0.794 1.297

15 IS 56.574 0.433 0.658 1.089

16 HOMO 56.181 0.474 0.688 1.150

17 GRF 52.591 1.082 1.040 1.629

18 DOG 53.460 0.886 0.941 1.595

19 DCT 52.716 1.052 1.026 1.744

20 ASR 49.419 2.248 1.499 2.543

21 AS 58.199 0.298 0.546 0.931

22 ANL 52.418 1.127 1.061 1.802

(1)

(2)

Engineering Letters, 24:1, EL_24_1_08

(Advance online publication: 29 February 2016)

______________________________________________________________________________________

where

, and

The following outlines the algorithm implementation

procedures of the proposed DCT using JPEG coding of a

still image:

1. Load the images of same sizes from a database

2. Implement a 2-dimensional DCT

3. Initialize the matrix to zeros

4. Perform a block DCT for a given image on 8 x 8 blocks

5. Apply JPEG default normalization matrix (QM) to

normalize the DCT transformed coefficients

6. Scale the DCT transformed image by maximum

coefficient value in the transformed image (i.e. a scale

factor, QP)

7. Perform an integer conversion of the DCT transformed

image

8. Select the number of DCT coefficients 1 through 64 for

the image to be compressed, this is called Quantization

9. Perform ordering of the transformed coefficients by zig-

zag scan using Matlab matrix indexing power

10. Construct Run-Level coding by pairing the Zig-Zaged

single column

11. Save the required bit streams as feature vectors during

enrollment, then compare the saved features with new

features during testing and

12. Measure and evaluate the performance (verification and

identification rates) of each algorithm by using nearest

neighbor classifier.

After computing the similarity matrices the results are

evaluated graphically and some numerical results are

displayed. The proposed framework is demonstrated in Fig.

6.

Fig. 6. Framework of the ASDCT for enrollment and verification

(0,3)(0,5)

F.

vectors

66 5 0 -2 0 -1 0 0

9 1 -1 2 0 1 0 0

14 1 -1 0 -1 0 0 0

3 -1 -1 -1 0 0 0 0

2 -1 0 0 0 0 0 0

0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0

Q

T

rain

ing

Sch

eme

T

esti

ng

Sch

eme

DCT

DCT

Q

Matching using

NNC

66 5 0 -2 0 -1 0 0

9 1 -1 2 0 1 0 0

14 1 -1 0 -1 0 0 0

3 -1 -1 -1 0 0 0 0

2 -1 0 0 0 0 0 0

0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0

Preprocessing +AS

Preprocessing +AS Probe Image

Probe Image

Accept/Reject

Face gallery

Engineering Letters, 24:1, EL_24_1_08

(Advance online publication: 29 February 2016)

______________________________________________________________________________________

Fig. 7. Sample of results from ORL (first row), Yale (second row) and

extended Yale-B (third row) showing gray scaled images, normalized

images using AS and DCT matrices across the column.

Fig. 8 and Fig. 9 show performance curves of the

ASDCT on ORL database. The performance is

demonstrated based on aggregate statistics and relative

ordering of match scores. The new method achieved

highest recognition accuracy.

Fig. 8. ROC curve of ASDCT with MAHCOS distance on ORL face

database

Fig. 9. CMC curve of ASDCT with MAHCOS distance on ORL face

database

The ASDCT algorithm proved high performance over

the other considered algorithms with verification rate of

86.00% at 0.1%FAR using Yale database. The latter

performed better using this database. The Fisherface

(LDA) method appears to be the best after the new method

in terms of suppressing the illuminations, with 74.44% at

0.1%FAR. Fig. 10 shows the representations of the entire

results generated from Yale face database.

Obviously, the results are expected to vary across the

databases. Receiver operating characteristic (ROC) and

cumulative match characteristic (CMC) curves are

generated to describe the results of each method. The

verification rates were presented at different FAR in the

ROC curves; while CMC curves depicted the rank

recognition rates.

10-3

10-2

10-1

100

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

False Accept RateV

erification R

ate

ASDCT+MAHCOS

The following popular appearance based algorithms

were exploited: PCA, Kernel Principal Component

Analysis (KPCA), Linear Discriminant Analysis (LDA)

and Kernel Fisher Analysis (KFA) for performance

comparison with the proposed method ASDCT. The

performance effectiveness of each algorithm was

measured and evaluated using Mahalanobis Cosine

(MAHCOS) similarity. The verification and

identification rates were recorded.

IV. RESULTS AND DISCUSSION

In order to observe the behavior of recognition

process, different algorithms have to be tested on some

selected databases. And extensive experiments have to

be conducted and evaluated by varying False

Acceptance Rate (FAR). Then the verification results

should be observed normally at 0.01, 0.1 and 1.0

percentages of the FAR. Fig. 7 shows the DCT matrix

obtained from the normalized images using the AS. The

normalizing effect of this technique can be visually

observed within certain limit. With poor illuminations it

is difficult for all algorithms and classifiers to recognize

the faces, since illuminations appeared in images as

noise. However, recognition with DCT can be greatly

improved by applying the AS for compensating

illumination in uncontrolled enviroments.

100

101

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Rank

Recognitio

n R

ate

ASDCT+MAHCOS

Engineering Letters, 24:1, EL_24_1_08

(Advance online publication: 29 February 2016)

______________________________________________________________________________________

Fig. 10. Performance comparison of ASDCT and other methods on Yale

face database

Similarly, using the ORL database, all the algorithms

performed better. And there is a significant improvement in

the Fisherface method from 74.44% to 78.34% at 0.1%

FAR. In this regard, the new technique outperformed the

others with up to 93.40% at 0.1% FAR, while it remains

better with other databases. Fig. 11 shows the comparison of

the results generated from ORL face database.

Fig. 11. Performance comparison of ASDCT and other methods on ORL

face database

All the algorithms observed a significant performance

reduction using the extended Yale-B, except the new method

with up to 83.75% at 0.1% FAR. This is because most of the

images in this database were poorly illuminated and poor

illumination hinders recognition accuracy. Furthermore,

removing some principal components does improve the

performance of the Eigenface method in the presence of

lighting variation. However, it does not achieve minimal

error rates compared to other methods described here. The

KPCA exhibited lowest performance with only 30.57% at

FAR=0.1%. Fig. 12 illustrates the results conducted on

extended Yale face database B.

Fig. 12. Performance comparison of ASDCT and other methods on

extended Yale face database B

The summary of the results are presented in Table II. It is

clearly evidence that the performance of the proposed

method is much higher than the other techniques reported on

all the databases.

TABLE II

PERFORMANCE COMPARISON OF ASDCT AND OTHER TECHNIQUES

CONSIDERED

For thorough candidate’s authentication analysis, other

performance metrics were equally generated through the

identification experiment. Equal error rate (EER), minimal

half total error rate (MHTER) and rank one recognition rates

representing the different techniques using ORL database are

recorded in Table III. The ASDCT reported the lowest EER

of 2.73%, and MHTER of 1.98%.

V

erif

icat

ion

Rat

e (%

)

V

erif

icat

ion

Rat

e (%

)

V

erif

icat

ion

Rat

e (%

)

DATABASE METHOD

VERIFICATION RATE (%)

0.01% FAR 0.1% FAR 1% FAR

Yale

PCA 45.31 66.52 83.99

LDA 64.34 74.44 90.10

KFA 52.16 63.63 76.88

KPCA 25.45 49.45 68.82

ASDCT 80.00 86.00 95.95

ORL

PCA 46.01 67.89 85.97

LDA 65.34 78.34 92.01

KFA 52.02 61.84 82.43

KPCA 28.00 52.03 70.10

ASDCT 82.80 93.40 97.70

Extended

Yale-B

PCA 27.90 47.01 61.50

LDA 55.55 64.21 67.09

KFA 38.48 44.68 68.89

KPCA 21.30 30.57 58.87

ASDCT 66.00 83.75 84.57

Engineering Letters, 24:1, EL_24_1_08

(Advance online publication: 29 February 2016)

______________________________________________________________________________________

TABLE III

RESULTS OF ERROR RATES, IDENTIFICATION RATES AND RANK ONE

RECOGNITION RATES ON THE EVALUATION DATA

METHODS (%) MHTER (%) EER (%) RANK (%)

PCA 4.53 4.96 65.85

KPCA 7.90 8.99 50.3

LDA 3.96 4.32 87.77

KFA 7.01 7.35 84.78

ASDCT 1.98 2.73 93.5

From the basis of evaluation of the 22 IN techniques, and

emperical analysis of the new method, it is deduced that

applying AS in the initial stage enhances recognition

accuracy of the DCT significantly. Moreover, several works

using DCT and other popular feature extraction techniques

were reported for further comparison with the proposed

method. The emphasis was given on ORL because it is the

most commonly used database especially on face recognition

using the DCT that outperformed other databases.

The research investigated one of the challenges of image

acquisition process, and proposed an improved approach of

a single DCT. The performance comparison of the new

method with few hybrid systems is presented in Table IV. In

this regard, a competitive advantage can be observed.

Therefore, the effect of illumination is addressed.

TABLE IV

COMPARISON OF RECOGNITION RATES OF THE PROPOSED ALGORITHM WITH

OTHER TECHNIQUES USING DIFFERENT DATABASES

METHOD DATABASE RECOGNITION RATE

(%)

PCA [30] ORL 90

PCA [31] Yale 81.13

MLP [27] ORL 77.20

DCT+MLP [27] ORL 92.90

HMM [32] ORL 87.00

DCT [10] CIM 84.58

KLT [10] CIM 77.57

Hexagonal +

DCT+MLP [33]

Yale 92.77

DCT [34] BioID 92.50

DCT+ DT-WT [35] Yale-B 88.30

DCT+ DT-WT [35] ORL 83.30

DLDA [36] Yale 82.14

D-DCT [37] Yale 90.14

ODCT [38] Yale 92.12

CCODCT [38] Yale 91.37

DFT+DCT [39] Color FERET 80.23

DCT [40] ORL 88.75

Proposed method ORL 93.40

The justifiable reason of carrying out the study is; some

feature extraction algorithms are more sensitive to

illumination than others, and no algorithm is however

insensitive to illumination variations. Furthermore, each of

the IN techniques can work better with some feature

extractors than the others. Therefore, recognition using AS

and few selected coefficients of the DCT reported additional

performance over the traditional face recognition techniques.

The AS technique estimates luminance function using

anisotropic smoothing and supports decorrelation of data in

this experiment. While the proposed method preserves

object boundary effectively without degrading the quality of

images. Fig. 13 demonstrates the effect of AS on a poor

illuminated image after the reconstruction at various

compression ratios of 70%, 50%, 30%, and 10%. The

advantage of the illumination compensation can be visually

compared with unfiltered reconstructed images at the same

compression ratio. Measurement cost and classification

accuracy are two major reasons of minimizing the dimension

of features [41]. This is achieved by discarding some of the

extracted features that possess low discriminating ability.

The selection of 19 coefficients in this research is

approximately 70% of the compression ratio.

Fig. 13. Effect of AS (second row) on an image (first row): (a) original

image, (b) image at 70% compression ratio, (c) image at 50% compression

ratio, (d) image at 30% compression ratio, and (e) image at 10%

compression ratio

Hence, anisotropic diffusion IN technique based Discrete

Cosine Transform’ (ASDCT) is introduced. The research

mitigates the effect of illumination variations on face images

using the most suitable illumination invariant, prior to

features selection using the DCT.

The major contributions of the proposed algorithm are: 1)

it promotes decorrelation of DCT sufficiently and

supersedes the appearance based techniques considered. 2) It

compensates the effects of illumination and creates almost

unvarying input images. The limitation of this study is we

can only assure the use of AS for efficient recognition using

DCT algorithm. Since the discovery of the most suitable

illumination compensator was carried out through JPEG

standard method, which is one of the typical image

compression techniques that use the DCT. Therefore, to

determine a preprocessor that works best with any feature

extractor, the property of the extractor needs to reflect the

testing criterion.

V. DISADVANTAGE OF TRUNCATING DCT

COEFFICIENTS

In all spectral computations, signal is truncated before the

discretization by multiplying the original signal say by

a rectangular window say , the resulted spectrum of

the truncated signal equals the convolution of and the

spectrum of the untruncated signal. Generally, truncation

introduces ripples, leakage or smearing into a spectrum [42].

The spectrum of a DCT signal is bounded at discrete

frequencies, and the cosinusoidal frequencies are disturbed

in an attempt to put a rectangular window. Although, the

coefficients of the DCT are truncated to reduce the

(a) (b) (c) (d) (e)

Original image

Original image

(a) (b) (c) (d) (e)

Original

Original

Engineering Letters, 24:1, EL_24_1_08

(Advance online publication: 29 February 2016)

______________________________________________________________________________________

complexity and redundancy, but at the expense of errors and

noise induction into the system. The quantization is carried

out by varying step sizes of DCT coefficients through fixing

a default quantization step size for each coefficient based on

visual sensitivity to different frequencies.

Therefore, the quantized signal is number of retained

coefficients out of the 64 coefficients of the matrix. The

larger the size of , the better the quality of reconstructed

image. The limitation of DCT is always known since

truncating higher spectral coefficients blurs the images

especially wherever the details are high. And coarse

quantization on some low spectral coefficients introduces

graininess in the smooth portions of the images.

Additionally, serious blocking artifacts are introduced at the

block boundaries since each block is independently encoded,

often with a different encoding strategy and the extent of

quantization [43]. And the size of the subset of DCT

coefficients retained as a feature vector may not be large

enough for achieving an accurate reconstruction of the input

image [10]. One of the phenomena that may decrease this

effect is by applying an overlapped transform.

Therefore, for an efficient decorrelation of images using

the approach mentioned in this paper, 19 coefficients

including the DC components were extracted. But the

texture based illumination compensator was resulted to a

smoothed matrix and uniform coefficients magnitude.

Besides, the truncation affects the system and these effects

were studied in [27], using error metrics; PSNR and MSE.

Similarly, reduction of a number of features may reduce the

discriminating power as a result lower the accuracy of the

recognition system [41]. Notwithstanding, reconstruction

without error cannot be achieved with truncation of any of

the frequency components either high or low. For this

reason, an advanced feature extraction method as a way

forward for improving data compaction of the DCT is

needed without truncating any of its coefficient.

VI. CONCLUSIONS

Out of The 22 IN techniques in the literature, the AS is

proven as the most suitable preprocessor for face recognition

using the DCT. The research used few of the DCT

coefficients to increase classification accuracy for optimal

feature extraction. Hence, enhanced DCT technique is

developed for an efficient face recognition. The proposed

method improved the performance of face recognition

system under both controlled and uncontrolled illumination

conditions. It also outperformed many conventional

techniques, and can be tested on other biometric modalities

such as palmprint.

Further research may be conducted on the block-based

ASDCT with more databases. Finally, considering a

significant loss of recognition rates generally using the DCT

which is caused by necessitating truncation, we will also

explore a new feature extraction that optimizes the matrix

coefficients and increases enrollment of large users within a

small memory disk, and without truncation of any DCT

coefficient.

REFERENCES

[1] A. Ghandehari & R. Safabakhsh, “A comparison of Principal

Component Analysis and Adaptive Principal Component Extraction

for palmprint recognition”, IEEE, 2011.

[2] D. Bhattacharyya, R. Ranjan, F. A. Alisherov and M. Choi,

“Biometric authentication: A review”, International Journal of u-

and e- Service, Science and Technology, vol. 2, no. 3, pp. 13-27.

2009.

[3] H. Drira, B. B. Amor, M. Daoudi, and R. Slama, “3D Face

recognition under expressions, occlusions and pose variation, IEEE

Transactions on pattern analysis and machine intelligence, vol. 35,

no. 9, 2013.

[4] S. T. Gandhe, K. T. Talele, and A. G. Keskar, "Face recognition using

contour matching," IAENG International Journal of Computer

Science, vol. 35, no. 2, pp226-233, 2008.

[5] R. Hietmeyer, 2000. Biometric identification promises fast and

secure processing of airline passengers, The International Civil

Aviation Organization Journal, vol.55, no.9, pp. 10-11.

[6] G.Hemalatha, C.P. Sumathi, “A study of techniques for facial

detection and expression classification” International Journal of

Computer Science & Engineering Survey (IJCSES), vol.5, no.2,

2014. DOI : 10.5121/ijcses.2014.5203

[7] V. Radha, and N. Nallammal, "Neural network based face

recognition using RBFN classifier," Proceedings of the World

Congress on Engineering and Computer Science, vol. I, pp1-6,

2011.

[8] Z. Sufyanu, F. S. Mohamad and A. S. Ben-Musa, “Choice of

illumination normalization algorithm for preprocessing efficiency of

Discrete Cosine Transform”, International Journal of Applied

Engineering, vol. 10, no. 3 pp. 6341-6351, 2015.

[9] Face Recognition, Technical Report. National Science and

Technology Council. Web source:

www.biometrics.gov/Documents/facerec.pdf.

[10] Z. M. Hafed, M. D. Levine, “Face recognition using the Discrete

Cosine Transform”, International Journal of Computer Vision,

vol.43, no.3, pp. 167–188, 2001.

[11] N. Ahmed, T. Natarajan, K. Rao, Discrete Cosine Transform, IEEE

Trans. on Computers vol. 23, no. 1, pp. 90–93, 1974.

[12] K. Rao, P. Yip, Discrete Cosine Transform, Algorithms,

Advantages, Applications. Academic: New York, 1990.

[13] B. K. Ahmed. “DCT image compression by Run-Length and Shift

Coding techniques.” J. of university of Anbar for pure science,

vol.5, no.3, 2011. Web source: http://www.e-

bookspdf.org/download/text-compression-coding-techniques.html

[14] S. K. Dandpat, and S. Meher, New Technique for DCT-PCA Based

Face Recognition, 2011.

[15] C. Sanderson, and K. K. Paliwal, “Polynomial features for robust

face authentication”, Proceedings of international conference on

image processing, IEEE, vol. 3, pp. 997-1000, June 2002.

[16] Z. Sufyanu, F. S. Mohamad and A. A. Yusuf, “A new discrete cosine

transform on face recognition through histograms for an optimized

compression” Research Journal of Information Technology,

DOI:10.3923/rjit.2015

[17] A. A. Fathima, S. Vasuhi, N. T. N. Babu, V.Vaidehi, and T. M.

Treesa, "Fusion framework for multimodal biometric person

authentication system," IAENG International Journal of Computer

Science, vol. 41, no. 1, pp18-31, 2014.

[18] C. Atupelage, and K. Harada, "Quantifying the visual content

perceived in DCT coefficients and PKI based semi-fragile

watermarking for visual content authentication of images,"

Advances in Electrical and Electronics Engineering-IAENG Special

Edition of the World Congress on Engineering and Computer

Science, pp275-286, 2008.

[19] S. Vikas, and J. P. Gupta, "Collusion attack resistant watermarking

scheme for colored images using DCT," IAENG International

Journal of Computer Science, vol. 34, no. 2, pp171-177, 2007.

[20] A. Reza, B. Azad, and S. Jamali. "Human face recognition with high

accuracy using Diverse Feature Set and mixture of SVMs."

International Journal of Modern Education and Computer Science

(IJMECS). DOI: 10.5815/ijmecs.2014.08.08, vol. 6, no. 8, pp. 66-

74, 2014.

[21] F. Juefei-Xu, and M. Savvides. "Subspace based discrete transform

encoded local binary patterns representations for robust periocular

matching on NIST’s face recognition grand challenge." In IEEE

transactions on image processing, vol. 1, no. 1, pp. 1-16, 2014.

[22] M. Lee, and C. H. Park. "An efficient image normalization method for

face recognition under varying illuminations." In proceedings of the

Engineering Letters, 24:1, EL_24_1_08

(Advance online publication: 29 February 2016)

______________________________________________________________________________________

1st ACM international conference on multimedia information

retrieval, pp. 128-133, 2008.

[23] S. Liao, D. Yi, Z. Lei, R. Qin, and S. Z. Li, "Heterogeneous face

recognition from local structures of normalized

appearance", Advances in biometrics. springer berlin heidelberg, pp.

209-218, 2009.

[24] X. Tan and B. Triggs. "Enhanced local texture feature sets for face

recognition under difficult lighting conditions." In IEEE transactions

on image processing, vol. 19, no. 6, pp. 1635-1650, 2010.

[25] S. Du and R. K. Ward. "Adaptive region-based image enhancement

method for robust face recognition under variable illumination

conditions", In IEEE transactions on circuits and systems for video

technology, vol. 20, no. 9, pp. 1165-1175, 2010.

[26] Z. Lai, D. Dai, C. Ren and K. Huang "Multi-layer surface albedo for

face recognition with reference images in bad lighting conditions." In

IEEE transactions on image processing, vol. 23, no. 11, pp. 4709-

4723, 2014.

[27] Z. Pan, H. Bolouri, High Speed Face Recognition based on Discrete

Cosine Transforms and Neural Networks, 1999.

[28] R. Gross, V. Brajovic. “An image preprocessing algorithm for

illumination in-variant face recognition”, Proc. of theyu 4th

International conference on audio and video-based biometric

personal authentication, July.

[29] K.C. Lee and J. Ho and D. Kriegman, "Acquiring linear subspaces for

face recognition under variable lighting ", IEEE Trans. Pattern anal.

mach. intelligence, vol. 27, no. 5, pp. 684-698, 2005.

[30] F. Samaria, Face recognition using Hidden Markov Models. PhD

thesis, Trinity College, University of Cambridge, Cambridge, 1994.

[31] R, Huang, V. Pavlovic, and D.N. Metaxas, “A hybrid face

recognition method using Markov Random Fields”, ICPR04, pp. 157-

160, 2004.

[32] F. Samaria and A. Harter, “Parameterisation of a stochastic model for

human face identification,” in proceedings of 2nd IEEE Workshop on

applications of computer vision, (Sarasota Florida, USA), Dec. 1994.

[33] M. Azam, M. A. Anjum and M. Y. Javed, “Discrete Cosine Transform

based face recognition in hexagonal images,” IEEE, vol. 2, 2010.

[34] A. R. Chadha, P. P. Vaidya, and M. M. Roja, “Face recognition using

Discrete Cosine Transform for global and local features,”

Proceedings of the 2011 international conference on recent

advancements in electrical, electronics and control engineering

(IConRAEeCE) IEEE Xplore: CFP1153R-ART.

[35] K. Ramesha, and K. B. Raja, “Dual transform based feature extraction

for face recognition”, International Journal of Computer Science

Issues, vol. 8, Issue 5, no. 2, Sept. 2011.

[36] H. Yu and J. Yang, “A direct LDA algorithm for high-dimensional

data with application to face recognition”, Pattern Recognition, vol.

34, no. 10, pp. 2067-2070, 2001.

[37] X. Y. Jing and D. Zhang, “A face and palmprint recognition approach

based on discriminant DCT feature extraction”, IEEE Trans. on Syst.

Man Cybern B, vol. 34, no. 6, pp. 2405-2415, 2004.

[38] P. Cui, and R. B. Zhang, “A semi-supervised coefficient select

method for face recognition”, Journal of Harbin Engineering

University, vol. 33, no. 7, pp. 855-861, 2012.

[39] G.M. Deepa, R. Keerthi, N. Meghana, and K. Manikantan, “Face

recognition using spectrum-based feature extraction”, Applied Soft

Computing Journal, vol. 12, no. 9, pp. 2913-2923, 2012.

[40] P. Gupta and N. V. George, “An improved face recognition scheme

using transform domain features,” International Conference on

Signal Processing and Integrated Networks, 2014.

[41] J. k. Anil, R. P. W. Duin, and J. Mao. "Statistical pattern recognition:

A review." Pattern Analysis and Machine Intelligence, IEEE

Transactions on 22.1 (2000): pp. 4-37.

[42] C. Chi-Tsong, Digital signal processing: spectral computation and

filter design. Oxford University Press, 2001, pp. 101-103.

[43] Su, Po-Chyi, Houng-Jyh M. Wang, and C-C. Jay Kuo. "Digital

watermarking on EBCOT compressed images." SPIE's International

Symposium on Optical Science, Engineering, and Instrumentation.

International Society for Optics and Photonics, 1999.

Engineering Letters, 24:1, EL_24_1_08

(Advance online publication: 29 February 2016)

______________________________________________________________________________________